How words can and cannot be learned by observation
Tamara Nicol Medinaa,b,1, Jesse Snedekerc,1, John C. Trueswella,b,1, and Lila R. Gleitmana,b,1
a
Department of Psychology and bInstitute for Research in Cognitive Science, University of Pennsylvania, Philadelphia, PA 19104; and cDepartment of
Psychology, Harvard University, Cambridge, MA 02138
Three experiments explored how words are learned from hearing
them across contexts. Adults watched 40-s videotaped vignettes
of parents uttering target words (in sentences) to their infants.
Videos were muted except for a beep or nonsense word inserted
where each “mystery word” was uttered. Participants were to
identify the word. Exp. 1 demonstrated that most (90%) of these
natural learning instances are quite uninformative, whereas
a small minority (7%) are highly informative, as indexed by participants’ identification accuracy. Preschoolers showed similar
information sensitivity in a shorter experimental version. Two further experiments explored how cross-situational information
helps, by manipulating the serial ordering of highly informative
vignettes in five contexts. Response patterns revealed a learning
procedure in which only a single meaning is hypothesized and
retained across learning instances, unless disconfirmed. Neither
alternative hypothesized meanings nor details of past learning
situations were retained. These findings challenge current models
of cross-situational learning which assert that multiple meaning
hypotheses are stored and cross-tabulated via statistical procedures. Learners appear to use a one-trial “fast-mapping” procedure, even under conditions of referential uncertainty.
acquisition
| induction | language | vocabulary
F
undamental for each child entering the human community is
the acquisition of word meanings: discovering which language
sounds map onto which interpretations. Because these mappings
are arbitrary and vary cross-linguistically, growing a vocabulary
poses a classic learning problem for humans, both infant learners
of a first language and second-language learners who must replace the original mappings with a new set. This experiencedependent learning problem for humans contrasts with animal
communication systems in which the interpretations of speciesspecific barks, chirps, and growls are largely given for free by
nature. The present article provides experimental evidence concerning the primitive initial procedure by which humans acquire
vocabulary items.
A common assumption is that form-to-meaning mappings are
discovered in a process mediated by observation of extralinguistic events: The learner matches recurrent speech events to recurrent aspects of the observed world. For example, when an
English speaker says “dog” or a French speaker says “chien,” there
is likely to be a co-occurring dog sighting. Young children often
acquire a word’s meaning after a single such exposure to its use in
context (1), particularly if there is strong pragmatic support (2) or
a restrictive syntactic environment (3). The sheer size of the average vocabulary at age 6 y [estimated at 6,000–8,000 words (1)]
suggests that this “fast mapping” of a sound segment onto its interpretation must happen very often, as is also attested in many
laboratory studies (4, 5).
Yet the world of words and their contexts is enormously
complex. Few words are taught systematically, even in middleclass environments with their blocks and picture books. Rather,
in most instances, the situations of word use arise adventitiously
as adults interact socially with novices. Words are heard buried
inside multiword utterances and in situations that vary in almost
endless ways—the bath, the zoo, the supermarket—so that usually a listener could not be warranted in selecting a unique interpretation for a new item. For example, in the fairly typical
setting of Fig. 1A, there are hundreds of objects, happenings,
properties, and relations that might be picked out as a match
www.pnas.org/cgi/doi/10.1073/pnas.1105040108
when the adult utters some new word: one of the objects (shoe,
pacifier), qualities (black, large) or parts (lace, sole) of the object, the child’s own actions (bump, look), and so forth. Worse,
the match could be to any two or more of these categories, e.g.,
“black shoe” or “Mommy’s shoe.” Also, speakers often allude to
absent objects and events (“Let’s go to the zoo,” “Remember
when we lost your ball?”) or ineffable properties (“Be nice
to your little brother,” “Time for your nap.”). A word once heard
may not be encountered again for days or weeks, during which
hundreds of other words intervene. At the other extreme, those
things most consistently present—the ceiling, an eye, breathing—
are often least likely to be topics of conversation; rather, their
omnipresence diminishes their salience as conversational focus.
Thus, to say that a word is learned “by” observation must be true
in some way, but so saying leaves much to the imagination about
actual procedures.
A venerable assumption that addresses some of these problems is that learning happens over several encounters: Word
meaning can be acquired probabilistically and incrementally by
accumulating evidence across several situations in which the same
phonological segment occurs. Hearing the same sound under new
circumstances can eliminate false conjectures or add detail to
underspecified conjectures. Many researchers have tested this
hypothesis with laboratory simulations (6–10).* In these experiments, pictures of objects appear on a neutral background (Fig.
1B). Each image is assigned a nonsense word, with the participants’ task being to discover these mappings. To reproduce
uncertainty, several images are shown simultaneously on each
trial, along with an equal number of auditorily presented nonsense words. Over trials, participants could track co-occurrences
to determine which sound-image pairs are reliably copresent.
Adults and children are quite adept at solving such laboratory
problems, performing above chance after several such exposures.
At least two major questions, discussed below, arise about
the relevance of such laboratory demonstrations to real vocabulary acquisition.
Contextual Uncertainty. As already suggested, in ordinary conversational settings hundreds of objects are in view, rather than
three or four, many of them plausible referents. A theory stipulating that all such context-appropriate alternatives are stored,
awaiting cross-situational disambiguation, grows less appealing
as their number increases. That is, constraints on memory for
past contexts may be a more severely limiting factor in natural
learning than appears in stripped down laboratory demonstrations. Far more troublesomely, simulating “context” using
Author contributions: T.N.M., J.S., J.C.T., and L.R.G. designed research; T.N.M. and J.S.
performed research; T.N.M., J.S., J.C.T., and L.R.G. analyzed data; and T.N.M., J.S., J.C.T.,
and L.R.G. wrote the paper.
The authors declare no conflict of interest.
1
To whom correspondence may be addressed. E-mail: gleitman@psych.upenn.edu, medinatn@sas.upenn.edu, trueswell@psych.upenn.edu, or snedeker@wjh.harvard.edu.
This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10.
1073/pnas.1105040108/-/DCSupplemental.
*The kinds of words studied here (basic-level whole-object terms that surface as nouns in
most languages) constitute only a small proportion of words known and used by 3- and
4-y-olds. Nevertheless, these items predominate in the vocabularies of infants (11) and
novice older learners of both first and second languages (12). These words constitute the
enabling basis for more advanced vocabulary growth (1, 3) and require the least internal
linguistic support to acquire (13, 14). Abstract words such as time or think have no
identifiable images and are acquired via procedures that make crucial use of their
syntactic and discourse contexts (15).
PNAS Early Edition | 1 of 6
PSYCHOLOGICAL AND
COGNITIVE SCIENCES
Contributed by Lila R. Gleitman, April 2, 2011 (sent for review November 23, 2010)
performance after several image-to-sound pairings), an outcome
that (depending on the dataset) might be generated by all-or-none
models as well as statistical-accrual ones. [One study (10) did collect trial-by-trial conjectures, but did not consider the possible
strategy of remembering a single hypothesis across trials.] In
the experiments reported below we also examine how learners’
interpretations evolve across successive contextually uncertain
encounters.
Fig. 1. (A) A plausible word learning environment for the word shoe. (B)
The simulated word-learning environment for shoe found in most crosssituational word-learning experiments.
repeated identical images seriously distorts reality, for learners
hear the word dog not always and only in the presence of some
dog (or dog-image) frozen in time and place, but in the presence
of different dogs doing different things—and sometimes in the
absence of any dog . Hence, even with perfect memory, one may
never settle on a single correct meaning as exposures accumulate. Because words are uttered from time to time absent their
referents, the number of available hypotheses could actually increase across contexts, posing a principled conundrum for statistical learning machinery. A collateral possibility is that
learners may have implicit means for distinguishing between
more and less useful contexts, discarding some input without its
entering into the search for meaning (16). Our first aim in the
present experiments will be to understand the extent to which
statistical-accrual procedures, demonstrable for small fixed sets
of word-image pairings, scale up to the multiply ambiguous
contexts in which vocabulary is acquired.
Evaluating the Models. Although the models usually proposed for
laboratory word-learning findings are information-accrual based,
their major computational assumptions have not been tested behaviorally, but simply taken for granted. In outline, these schemes
propose that for each learning instance, all concepts known to the
learner are cross-tabulated with the word. If the situation is interpreted as supporting certain pairings, those associations are
strengthened (e.g., hearing “moop” in the presence of Fig. 1B leads
to simultaneous strengthening of the moop-shoe and moop-chair
associations, among others). Later encounters with moop in other
contexts lead to additional associative updates. As this theory
implies, even when each learning instance is low in informativity
(because any of the objects on the screen could be the referent of
moop), learners gradually reach consensus by tracking and
retaining all observed pairings, with the strongest associate chosen
at the end. However, no time-course data are provided for learning
across instances that might validate these assumptions. Rather, the
only measure reported is final response accuracy (above-chance
2 of 6 | www.pnas.org/cgi/doi/10.1073/pnas.1105040108
Experimental Findings
In three experiments, participants guessed word meanings by observing 40-s videos (“vignettes”) of naturally occurring contexts
(the meal, the play-room, and so forth) in which parents uttered
these words to their infants aged 12 to 15 mo. For example, one
vignette showed a mother opening a bag of toys while saying,
“Who else is in the bag?” to her child. Participants watched the
vignettes with the audio muted. They heard only an identifying
signal (either a beeping noise or a nonsense word, depending on
the experiment) occurring exactly when the mother had (really)
said, e.g., “bag.” Because new words are rarely introduced in an
uninterrupted sequence (“doggie, doggie, doggie. . .”) but rather
intermixed with other words, the vignettes for each target word
were distributed among all of the others during the experiments.
Given that these vignettes reflected a random sample of parent–
child interactions, the vignettes and their presentation conditions
reproduce critical properties of real environments in which early
word learning happens.
The vignettes were muted for two reasons. First, this replicates
the conditions of the experiments just reviewed, which present
word-image pairs without sentence context. Second, infants
learning their first words do not know the meanings of other
words in the utterance and thus cannot use these to determine
meaning; by removing this information we more accurately
model the earliest stages of acquisition, the point at which statistical accrual is argued to play its greatest role (3, 11). The
experiments used these vignette “inputs” to probe the learning
procedure for early vocabulary systematically, in terms of three
questions: (i) When do naturally occurring scenes offer information precise or relevant enough to trigger successful word
learning? (ii) How do repeated exposures to the same word in
different contexts channel and enhance this process? (iii) How
does memory influence operation of the learning machinery?
Exp. 1: Vignette Informativity. The contexts in which words are
uttered vary in informativity, ranging from cases where a parent
may gaze fixedly at an object the infant is holding as they name it
(“This is a horse”) to cases where no horse is even in sight (“Let’s
go see the horses at the farm”). Here we estimated the frequency
of such more and less informative contexts.
Thirty-seven adults watched randomly selected vignettes containing the 24 most frequently occurring nouns and 24 most
frequently occurring verbs in the video corpus developed by the
author J.S. There were 288 vignettes total, and groups of participants viewed different subsets: two vignettes for each of 48
words. The occurrence of each such “mystery word” was identified by a beep that was identical within and across items, so there
was no opportunity for cross-situational learning. The question
was which of these vignettes were sufficiently transparent that
observers could guess the speaker’s beeped-out word. A shorter
version of this experiment was run with 12 3- to 5-y-olds, who
watched one video for each of eight target words (SI Text).
Consistent with our prior findings (14), participants were poor at
guessing the intended referent of even highly frequent terms when
shown a single instance. Only for 7% of these vignettes did adult
participants achieve accuracy scores at or above 50% correct; we
classified these vignettes as High Informative (HI). In contrast,
90% of vignettes resulted in accuracy scores below 33%, thus
categorized as Low Informative (LI). It is of some interest that
none of the obtained HI vignettes were for verbs; all of them were
for common nouns labeling categories of whole objects (e.g., bag,
ball, horse) that are very frequent in the speech of parents to
children and appear in the vocabularies of 1- and 2-y-olds worldMedina et al.
Exp. 2: Learning with Cross-Situational Information. We now asked
how providing several contexts for new words affects learning.
We selected concrete whole-object items and set the ratio of HI
to LI vignettes at 1:4, to approximate their actual rate of appearance in mother-infant speech as ascertained from Exp. 1.
New participants identified 12 such “mystery words” by observing
five contexts for each. Each word was assigned a nonsense version (e.g., vash), replacing the beeps of Exp. 1 so that participants
could determine which vignettes went with which item, even
though they were intermixed in presentation order (e.g., a mipen
vignette, followed by a vash vignette, and so forth). Presence and
position of HI and LI vignettes was manipulated for four participant groups. One group saw five LI vignettes for each mystery
word (HI Absent). The other three groups saw four LI vignettes
and one HI vignette, but each in a different order: in one group
the HI vignette was the first for each target word (HI First), for
another it was third (HI Middle), and for the final group it was
fifth (HI Last). Participants recorded their guess after viewing
each vignette, allowing for later assessment of how knowledge
was evolving across interim information states. They also provided a confidence rating and a final guess for each word at the
end of the experiment.
Current statistical models evaluate each word-meaning hypothesis based on the properties of all past learning instances,
regardless of the order in which they are encountered (6–10),
holding numerous hypotheses in mind (with changing weights)
until some learning threshold is reached. This process predicts
that: (i) final conjectures will be indifferent to the sequential
position of HI vignettes; (ii) multiple meanings (i.e., traces of
past encounters) will be stored and tracked across vignettes; and
(iii) gradual learning will occur across LI vignettes. Here we test
a radically different, one-trial or “fast mapping” hypothesis, in
which (i) learners hypothesize a single meaning based on their
first encounter with a word; (ii) learners neither weight nor even
store back-up alternative meanings; and (iii) on later encounters,
learners attempt to retrieve this hypothesis from memory and
test it against a new context, updating it only if it is disconfirmed.
Thus, they do not accrue a “best” final hypothesis by comparing
multiple episodic memories of prior contexts or multiple semantic hypotheses.
The contrasting predictions of these two approaches are clear:
if learning is statistical, the position of HI in the sequence of
vignette presentations (the three experimental group conditions)
will produce perturbations in the learning curve early on but any
Medina et al.
such effect should be wiped out in the end, for by the fifth vignette all participants in all experimental conditions will have
received all and only the same contexts and can compare among
them; multiple hypotheses should be present, especially on early
vignettes; and performance accuracy should roughly rise across
vignettes. However, if learners posit a single hypothesis (and no
alternatives) on the first encounter and then seek to confirm or
disconfirm it, accuracy on the first vignette should be a major
determinant of accuracy by the fifth vignette, there should be
little or no evidence in interim vignettes of maintaining back-up
hypotheses, and response accuracy need not rise across vignettes.
Results
Effect of Information Order. Fig. 2 plots results as a function of
presence and order of HI information. As predicted by the onetrial learning hypothesis, when participants received an informative first exposure (HI First) they tended to: guess correctly
(66% accuracy on vignette 1); stick to that guess across vignettes
(41% accuracy by the fifth vignette); offer this guess as “the
meaning of the word” at experiment’s end (37% accuracy on
the Final Guess); and record strong and rising cross-trial confidence in this answer (SI Text). Accuracy on the fifth vignette and
final conjectures was significantly higher in the HI First condition
than in third (HI Middle) or fifth (HI Last) conditions (Table 1a,
Effect of learning condition on the fifth vignette, and Table 1b,
Effect of learning condition on the final conjecture). Not only
did good information arriving late fail to lead to correct performance in the end (contra statistical learning predictions), but
a set of low-informative vignettes failed to lead to any cross-trial
improvement (HI Absent) (Table 1c, Effect of vignette number
on accuracy in HI Absent condition).
The consequences of initial correct guessing are most apparent in Fig. 3, which plots accuracy independent of whether input
was “good” (i.e., HI) or not. In either case, a major determinant
of accuracy on later vignettes is accuracy on the first. [Data from
vignette 5 and final conjectures of the HI Absent condition
are excluded in Fig. 3 because vignette 5 repeated vignette 1
(SI Text).] Symmetrically, performance is extremely poor after an
incorrect guess, even when that incorrect guess was made on
a HI vignette. This latter finding contradicts current cross-situational models, because it shows that participants had little to no
memory of alternative word-meaning hypotheses that they could
return to. This finding vindicates the reality of the issue mentioned in introductory remarks, namely the impossibility of
storing the scores (if not thousands) of objects, qualities, and so
forth, that any single observation embodies, and that might be
relevant to the meaning of a new word then heard. Had participants stored some plausible alternative hypotheses after guessing incorrectly on a HI vignette, one of the hypotheses would
likely have been the correct one, leading to better performance
Fig. 2. Mean proportion of correct responses for each learning instance
(vignette) and the final conjecture. Error bars = 95% confidence intervals,
Exp. 2.
PNAS Early Edition | 3 of 6
PSYCHOLOGICAL AND
COGNITIVE SCIENCES
wide (11). This pattern is consistent with previous demonstrations
(14) that this concrete (or “imageable”) component of the lexical
stock is the easiest kind to acquire from passive observation of
situational context. Sixteen of the 24 nouns examined were of this
type; when we restrict the pool of vignettes to these words (i.e.,
excluding verbs and abstract nouns), the ratio of HI to LI vignettes
is approximately 1:5.
We also assessed whether preschoolers were sensitive to the
same aspects of the observed events as adults, and they were:
Children were more accurate on HI vignettes (53%) than LI
vignettes (22%).
In sum, this experiment exhibited the difficulty of determining
the meaning of words based on a random sample of single observations, even when items are all highly frequent in parental
speech. Expectation of this uncertainty has always motivated
cross-situational approaches to word learning. In addition to
providing some documentation of the magnitude of uncertainty in
parent–child interactive settings, the results provided a set of
vignettes that was classified for informativity. These vignettes were
used in Exps. 2 and 3 to see how cross-situational opportunities
might aid learning. Importantly, Exp. 1 demonstrates that adults
and young children cull these vignettes for information in much
the same way, so that we could continue to use adults as foils for
our learning questions with some confidence that the results would
link plausibly to the case of child word learning. These findings
further document the accumulating evidence that child and adult
word learning share essential commonalities (7, 12, 15).
Table 1. Exps. 2 and 3: Results of multilevel logistic regressions of binary data
Model
a. Effect of learning condition on
accuracy of the fifth (Final) vignette.
(Baseline = HI First)
b. Effect of learning condition on
accuracy of final conjecture.
(Baseline = HI First)
c. Effect of vignette number (1–5) on
accuracy in HI Absent condition.
d. Effect of first vignette accuracy on
accuracy of vignette 2.
e. Effects of confirmation and accuracy
on repetition of response
f. Effect of delay on accuracy of the HI
vignette. HI Last.
Effect
Intercept
HI First vs. HI Middle
Hi First vs. HI Last
Intercept
HI First vs. HI Middle
HI First vs. HI Last
Intercept
Vignette
Intercept
Accuracy of V1
Intercept
Confirmation
Accuracy
Intercept
Delay
Estimate
−0.5941
−1.9654
−1.2564
−0.7621
−1.8978
−1.963
−2.7558
0.02487
−2.9295
2.0704
−2.7684
1.9742
2.0269
−1.5874
1.3611
SE
0.5719
0.4935
0.4099
0.4724
0.4571
0.4902
0.86297
0.08608
0.6467
0.2952
0.222
0.2298
0.2414
0.4281
0.4892
Wald-Z
−1.039
−3.983
−3.065
−1.613
−4.152
−4.004
−3.193
0.289
−4.530
7.013
−12.471
8.59
8.397
−3.708
2.782
P value
<
<
<
<
<
0.29893
0.00007***
0.00218**
0.10700
0.00003***
0.00006***
0.00141**
0.77263
0.00001***
0.00001***
0.00001***
0.00001***
0.00001***
0.00021***
0.0054**
Models computed in R, using crossed random effects for Subjects and Items. See SI Text for R-syntax. Except for the model, Effect of vignette number (1–5)
on accuracy in HI Absent condition, models were significantly better fits compared with simpler models which did not include the reported fixed effects [χ2
tests of change in −2 restricted log likelihood (20)]. More complex models with additional fixed effects or interactions did not reliably improve fit. The “Effect
of vignette number (1–5) on accuracy in HI Absent condition” model is shown for illustration, as it is not a better fit than the null model. In the “Effect of
first vignette accuracy on accuracy of vignette 2” model, models including Informativeness of the first vignette (HI vs. LI) did not improve the fit. ***P < 0.001,
**P < 0.01, *P < 0.05.
later compared with initially guessing incorrectly on a LI vignette
(Table 1d, Effect of first vignette accuracy on accuracy of vignette 2). Statistical accrual models, (e.g., ref. 6), store all
plausible hypothesized meanings for later consideration on the
next learning instance.
Confirmation effect. A single-conjecture learning procedure based
on variable input would seem likely to acquire a vocabulary replete with errors, with one child calling the shoes “shoes” but
another child calling them “wheels” or “dogs,” depending on the
vagaries of first encounters. However, such errors are vanishingly
rare, even in toddlers. If learners are not (as in associative
schemes) maintaining and cross-tabulating several hypotheses,
what is preventing such errors? A clue comes from a further
feature of Fig. 2. Rather than accuracy rising between vignettes 1
and 2, as would be expected if several hypotheses were being
evaluated across learning instances, accuracy actually drops, a so
far unexplained and surprising consequence of added contextual
information. This decrement over vignettes is true even in our
“best” case of HI data received on vignette 1, for which the accuracy of over 60% on vignette 1 falls to about 40% by vignette 5.
Part of this performance drop no doubt reflects participants’
failing on occasion to recall their initial hypothesis, for overall
the experiment imposes a considerable memory burden. Even
Fig. 3. Mean proportion of correct responses for each vignette and the
final conjecture, plotted as a function of being correct or incorrect on vignette 1, split by whether vignette 1 was High informative (HI) or low Informative (LI). Error bars = 95% confidence intervals, Exp. 2.
4 of 6 | www.pnas.org/cgi/doi/10.1073/pnas.1105040108
closer to the heart of the learning problem for vocabulary is
participants’ likely bemusement by the striking differences in
contexts across vignettes, an issue that cannot arise in repeated
image-to-label experimental studies. Overall, the rarity of context-invariance suggests an implicit reliability check on the initial
conjecture before it is consolidated in memory.
Exp. 1 results provided the basis for investigating this issue.
Participants in Exp. 1 provided a range of conjectures for each
vignette, viewed in isolation. We reasoned that subsequent
vignettes in Exp. 2 would provide confirming evidence for the
initial hypothesis (the guess on vignette 1) just in case that hypothesis was a member of the Exp. 1’s response set for that word.
That is, if a vignette in isolation brings to mind a shoe, even if
ever so slightly, enough to have drawn a shoe response from at
least one of the 12 Exp. 1 participants who observed this item,
this would be enough to confirm an initial hypothesis of shoe in
the cross-situationally exposed participants. Indeed, this is the
case. The rate at which a participant repeated a response on the
next learning instance was reliably higher when the next vignette
confirmed that previous response (0.45) than when it did not
(0.19): confirmation quadruples the odds of response repetition
(Table 1e, Effects of confirmation and accuracy on repetition of
response). This finding is true for both correct (0.53 vs. 0.37) and
incorrect guesses (0.37 vs. 0.08), although getting confirmation of
an incorrect guess is exceedingly rare (only 133 of 1,536 incorrect
responses were confirmed by our criterion).
Taken together, these findings suggest that participants seek
supportive evidence in later encounters with a word; for correct
conjectures, confirmation is usually found. Thus there is a oneand-a-bit learning procedure with a built-in mechanism for
looking before you leap to a permanent entry in the mental
lexicon. Finally (see again the predictions above for one-trial
learning, and Fig. 2), note that late (HI Middle) recipients of
good information do not maintain it, perhaps because they are in
the grip of their first, false, hypothesis.
In sum, Exp. 2 findings document a learning procedure in
which learners form a single conjecture, retaining neither the
context that engendered it nor alternative meaning hypotheses.
The criterion for retention of this conjecture is whether it is even
minimally plausible for the context in which one next hears the
word. A corollary is that a false hypothesis, once formed, blocks
the formation of new ones. A by-product of this first-is-best (or
“template”) procedure is disproportionate dependence on the
presence of useful information first. For this reason, receiving
a LI vignette first typically results in participants never settling on
any one meaning for that word (we will revise the term “never”
Medina et al.
Exp. 3: “Never” Isn’t “Forever.” Results thus far suggest that word
learning requires useful information to be present on a first encounter. But we know that so unforgiving a procedure can’t describe reality; failing to acquire cow the first time you heard it
cannot doom you to a lifelong cow-gap in your mental lexicon. So
there must be a way to interpret the notion of “first encounter”
that does not have this fatal implication. To explore this issue, we
reran the HI Last and HI Middle conditions of Exp. 2, except
now participants had a 1- to 3-d break before encountering their
first HI vignette. Such a delay might prevent interference from
earlier LI vignettes via ordinary processes of forgetting.
Results
Indeed, accuracy on a HI vignette in fifth/final position (HI Last)
was significantly improved when it appeared first on a new day
(0.47) compared with performance in Exp. 2 (0.22) (P < 0.01)
(Table 1f, Effect of delay on accuracy of the HI vignette HI Last).
Similarly, accuracy on a HI Middle vignette improved numerically
(0.45 vs. 0.39), as did final conjectures for both (HI Last: 0.18 vs.
0.10; HI Middle: 0.17 vs. 0.09), although these latter differences
were not significant. This pattern suggests the delay between
observations allowed participants to begin learning anew. Whatever
false conjectures were entertained in the first experimental session
were apparently forgotten, reducing interference in the second
session. Performance did still drop somewhat on final conjectures,
as in Exp. 2. Nevertheless, a long delay after poor learning instances
precludes false conjectures to a useful degree. In effect, “HI Middle” and “HI Last” have been converted to a new “HI First” by the
passage of time. So tomorrow is another day for one-trial word
learning, just as it is for many other quotidian events.
Discussion
All attempts to understand vocabulary learning incorporate
a component that matches recurrent sounds with their situational contingencies. Although it is now recognized that concomitant linguistic and discourse information contributes heavily
to word learning (3), the observational component predominates
at the earliest stages and plays a continuing role throughout life.
Because contexts are so various and uncertain, it has generally
been believed that learners store several of them, at some point
retrieving and comparing among them to extract their recurrent
properties from among the variable ones. This compare-andcontrast supposition lies at the heart of recent cross-situational
learning experiments that attempt to simulate novice vocabulary
learning and model its character. The present findings challenge
this picture. Specifically, our results are incompatible with three
fundamental predictions of cross-situational learning: (i) indifference to input order, (ii) maintenance of multiple hypotheses, and (iii) improvement across trials.
Indifference to Input Order. If the learning procedure accrues information across exposures, all of it available for evaluating the
best fit at the end, it should not matter when in the input sequence the best information occurs. However, our results show
a massive effect of input order, with the first exposure being
decisive (and therefore performance being best in the HI First
condition), and essentially no learning after a false initial guess in
any condition (Figs. 2 and 3).
Medina et al.
Maintenance of Multiple Hypotheses. For the proposed learning
machinery to benefit from successive exposures, it must maintain
several options during early points in the input sequence. However,
our laboratory learners’ characteristic style was to store a single
hypothesis (Table 1d, Effect of first vignette accuracy on accuracy
of vignette 2), abandoning it only when disconfirmed (Table 1e,
Effects of confirmation and accuracy on repetition of response).
Improvement Across Trials. Another hallmark of cumulative
learning is increasing accuracy (though with perturbations)
across trials. However, in our studies, accuracy remains flat or
falls across vignette trials (Fig. 2). Only on a new day, when the
slate is wiped clean, does performance rise again (Table 1f, Effect of delay on accuracy of the HI vignette HI Last).
Our findings, to the extent they can be linked to learning
in vivo, begin to suggest that observational word learning is
hardly “cross-situational” in the intended sense but is materially
an insightful and almost indefeasible process. This outcome
is consistent with several recent laboratory demonstrations of
young children acquiring new words from single encounters, and
then seamlessly generalizing their use (17); it is in accord, as well,
with the astonishing rate of child vocabulary growth.
We believe that statistical-accrual approaches to word learning
have seemed plausible to many commentators, in part because the
learner’s task has been informally envisaged as selecting among
a few alternative interpretations, more as depicted in Fig. 1B than
as in Fig. 1A. Indeed, statistical models have proven adequate for
properties of language for which the learner’s hypothesis space is
known to consist of a very small closed set of options: either heavily
constrained by nature, as in the machinery for discovering the
phonetic inventory of a natural language given the psychophysical
properties of speech perception (18), or experimentally constrained, as in recent word-boundary learning experimentation,
which vary the relative positioning of a small number of CV syllables (19). Similarly, the laboratory experiments we have reviewed, purporting to show that statistical-associative machinery
can account for how humans extract concept-word mappings, limit
the search space to a small set of fixed images that occur in temporal lockstep with an equally small set of nonsense sounds.
However, actual word learning does not occur in anything resembling this picture-book setting but rather in uncontrolled, everchanging environments in which each word occurs in different
sentences with different exemplars inside different events, and
only loosely intercalated with them in time and place. When the
hypothesis space is very large, the role of memory becomes much
more prominent.† Our findings cohere within a perspective in
which memory limitations not only limit interpretive choices,
but actually rescue the learning machinery from the pitfalls of
false choice.
More specifically, our findings suggest that word learners are in
one of three pertinent states during the acquisition of each word.
In the initial state A, there is no mapping at all. Upon hearing the
word in context, the machinery makes a (single) conjecture, thus
passing into an interim state B. Now the machinery seeks confirmation, i.e., a new context at least weakly consistent with the one
formed at A. If this step succeeds, the conjecture is further solidified as a confident hypothesis of the word’s meaning. The
crucial question is what happens when there is failure at B. What
we see in the laboratory (Figs. 2 and 3) is that learners shift to a
new conjecture, but this shift comes at some cost. Rather than
returning to a state of semantic innocence (A), learners enter
a memorial limbo (B), which leaves some residue of confusion
that interferes with subsequent learning. This confusion is eliminated after a considerable delay, whereupon the machinery
†
A trivializing alternative explanation is that our requirement that participants overtly
state and record their interim conjectures reduced memory for prior conjectures. We
investigated this issue separately in a manipulation in which participants responded
overtly only during the second half of the experiment. This process had no impact on
performance in the second half, compared with another set of participants who were
overtly responding throughout.
PNAS Early Edition | 5 of 6
PSYCHOLOGICAL AND
COGNITIVE SCIENCES
when we introduce Exp. 3). This learning procedure of conjecture coupled with verification is most clearly revealed by considering what happens across successive vignettes as a function of
HI position. Participants in the HI First condition have the advantage of being more likely to start out with a correct hypothesis
just because the first vignette they encounter is a highly informative one, and therefore they will be more likely to find confirming evidence in later learning instances compared with those
who began with an incorrect hypothesis. Almost by definition,
later vignettes contain information that is more likely to serve as
confirmation of a correct guess than an incorrect one (e.g.,
a later shoe/vash vignette is more likely to be an event that
involves another shoe than another elephant).
returns to its initial state (A) and can again form a “first” conjecture (Exp. 3).
Conclusions
The proliferating memory burden inherent in storing multiple
uncategorized contexts during the learning process suggests that
recorded memories for past situations cannot be the central
vehicle for constructing a lexicon. If it were, even more intractable problems arise with the evaluation procedure that the
learner must instigate at some appropriate time (say, when 5 or
10 or 50 instances of “wode” or “blicket” have been heard and
their associated situational contexts squirreled away “neutrally”
in memory). What in particular could be true of all these wodish
and blickettesque scenes? Prior findings from our laboratory
have suggested that learners don’t—as might be hoped—home
in on a narrower and narrower interpretation as a consequence
of such repeated encounters (e.g., “object” on the first encounter, “toy” on the second, “ball” by the third). Instead, learners do
the reverse: moving from least to more inclusive—that is, more
abstract—interpretations as trials and, therefore, differences in
observations increase. That is, one can always find “some” similarity among a set of disparate situations if one’s conjecture
becomes vague and general enough. In any set of human encounters, someone is almost always “looking” and there is just
about always some “thing” in view. If the idea is to parse out of
scenes that which is common to all of them, this increasing abstractness (cross-situational generalization of the bad kind) is
bound to happen. And this is exactly what participants do in
experimental conditions that encourage remembrance of things
past, namely massed trials in which exposures to new words
are not interspersed with other words, new or known (14). Yet
everything we know about vocabulary growth tells us that word
learning does not work this way; rather, the early vocabulary
is highly concrete, even where learner age varies (12). One-trial
learning presupposes strong restrictions on the hypotheses
that will even be entertained (e.g., ref. 18) and about conditions on the interpretation of what is being observed. Such
conditions of “relevance” or “newsworthiness” apply everywhere
in human social intercourse, not only in the special situation of
word learning.
Such an analysis of the problem would suggest that word
learners would be well advised not to consult specific memories
of past situations during word learning. Rather, a process in
which a single conjecture is stored, defeasible only if countermanded by a failure during the confirmation phase, is a more
reasonable approach, and one that is strongly supported by the
experimental findings reported here. Learners interrogate scenes
and analyze them for a plausible meaning, forming a single
conjecture that, unless explicitly and rapidly countermanded, can
last a lifetime. So conceived, learners acquire the meanings of
concrete terms “from” observation, but only when the shoe fits.
1. Carey S (1978) Linguistic Theory and Psychological Reality, eds Halle M, Bresnan J,
Miller GA (MIT Press, Cambridge), pp 264–293.
2. Baldwin DA (1993) Infants’ ability to consult the speaker for clues to word reference.
J Child Lang 20:395–418.
3. Gleitman LR, Cassidy K, Nappa R, Papafragou A, Trueswell JC (2005) Hard words. Lang
Learn Dev 1:23–64.
4. Dollaghan CA (1985) Child meets word: “Fast mapping” in preschool children. Child
Dev 58:1021–1034.
5. Heibeck TH, Markman EM (1987) Word learning in children: An examination of fast
mapping. Child Dev 58:1021–1034.
6. Yu C, Smith LB (2007) Rapid word learning under uncertainty via cross-situational
statistics. Psychol Sci 18:414–420.
7. Smith LB, Yu C (2008) Infants rapidly learn word-referent mappings via crosssituational statistics. Cognition 106:1558–1568.
8. Xu F, Tenenbaum JB (2007) Sensitivity to sampling in Bayesian word learning. Dev Sci
10:288–297.
9. Vouloumanos A (2008) Fine-grained sensitivity to statistical information in adult word
learning. Cognition 107:729–742.
10. Smith K, Smith ADM, Blythe RA (2010) Cross-situational learning: An experimental
study of word learning mechanisms. Cogn Sci 35:480–498.
6 of 6 | www.pnas.org/cgi/doi/10.1073/pnas.1105040108
Methods
Participants. Undergraduates from the University of Pennsylvania and Harvard University participated for course credit or $10/h: 37 participants in Exp.
1, 64 in Exp. 2, and 35 in Exp. 3.
Stimuli. Stimuli for Exp. 1 were 288 40-s muted videos (vignettes) of parent–
toddler (ages 12–15 mo) interactions in natural settings, of the 48 most frequent nouns and verbs (24 each) in the corpus. Vignettes were aligned so that
30 s into the video, the parent uttered the target word (at which a beep was
heard). Previous studies found this duration to be sufficient to understand the
gist at the moment the target word was uttered (13, 14). Exps. 2 and 3 target
stimuli were a subset of these vignettes (40 total) and were examples of
a parent uttering one of eight common nouns (five vignettes each of bag, ball,
book, horse, necklace, nose, phone, and shoe). A spoken nonsense word (e.g.,
“mipen”) was spliced into the silent audio instead of the beep, with the same
timing. Each nonsense word corresponded to a target word (e.g., mipen = bag;
vash = shoe), thereby permitting cross-situational word learning. SI Text
describes selection criteria for the 40 target vignettes, which were based on
the results of Exp. 1.
Design of Exps. 2 and 3. For the five vignettes of each target word, occurrence
of the HI vignette was manipulated relative to the four LI vignettes: it occurred First, Middle, Last, or was Absent. The four LI vignettes were placed in
a fixed random order relative to the position of the HI vignette. In the HI
Absent condition, the four LI vignettes were followed by a repetition of the
first LI vignette.
Presentation of target words was intermixed (distributed presentation)
such that each participant saw the first instance of each target word one after
the other, followed by all of the second instances, and so forth. In Exp. 2,
participants were randomly assigned to one of the four conditions (HI First,
Middle, Last, or Absent) and one of two stimulus orders (forward or reverse).
Exp. 3 was the same except that only HI Middle and HI Last lists were run.
Procedure of Exps. 2 and 3. See SI Text, including procedure for Exp. 1, which
was similar to Exps. 2 and 3. Participants were tested in small groups in
a room with a video projector. Participants were told they would watch
videos of parents interacting with their children, with parents uttering 1 of
12 common words. The sound would be absent but the “mystery word”
would be indicated by a corresponding spoken nonsense word, played at the
exact moment that the parent uttered the mystery word. Participants had to
“figure out which English words correspond to these twelve nonsense
words, given the scenes that you watch in these videos.” After each vignette,
participants wrote their guess and rated confidence from 1 (low) to 5 (high).
Participants could not review their previous responses. At the end, participants were given a list of the nonsense words and recorded a final conjecture and rating for each word.
Exp. 3 was identical to Exp. 2, except subjects participated in two sessions
separated by 1 to 3 d. During the first session, participants viewed the
vignettes leading up to but not including the HI vignettes. In the second
session they viewed the HI vignettes and any remaining LI vignettes.
ACKNOWLEDGMENTS. This work was funded by National Institutes of
Health Grant 1-R01-HD-37507.
11. Gentner D, Boroditsky L (2001) Language Acquisition and Conceptual Development,
eds Bowerman M, Levinson S (Cambridge University Press, Cambridge), pp 215–256.
12. Snedeker J, Geren J, Shafto CL (2007) Starting over: International adoption as
a natural experiment in language development. Psychol Sci 18:79–87.
13. Snedeker J, Gleitman LR (2004) Weaving a Lexicon, eds Hall DG, Waxman SR (MIT
Press, Cambridge), pp 257–293.
14. Gillette J, Gleitman H, Gleitman L, Lederer A (1999) Human simulations of vocabulary
learning. Cognition 73:135–176.
15. Papafragou A, Cassidy K, Gleitman L (2007) When we think about thinking: The
acquisition of belief verbs. Cognition 105:125–165.
16. Csibra G (2010) Recognizing communicative intentions in infancy. Mind Lang 25:
141–168.
17. Booth AE, Waxman SR (2002) Word learning is “smart”: Evidence that conceptual
information affects preschoolers’ extension of novel words. Cognition 84:B11–B22.
18. Liberman AM, Cooper FS, Shankweiler DP, Studdert-Kennedy M (1967) Perception of
the speech code. Psychol Rev 74:431–461.
19. Saffran JR, Aslin RN, Newport EL (1996) Statistical learning by 8-month-old infants.
Science 274:1926–1928.
20. Steiger JH, Shapiro A, Browne MW (1985) On the multivariate asymptotic distribution
of sequential chi-square statistics. Psychometrika 50:253–264.
Medina et al.